]> git.openfabrics.org - ~shefty/rdma-dev.git/blobdiff - net/ipv4/inet_connection_sock.c
inet: Allowing more than 64k connections and heavily optimize bind(0) time.
[~shefty/rdma-dev.git] / net / ipv4 / inet_connection_sock.c
index f26ab38680de00a5b77c303e34774a078022d499..df8e72f07478cb738bcc2a0e08472929af473462 100644 (file)
@@ -93,24 +93,40 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
        struct inet_bind_hashbucket *head;
        struct hlist_node *node;
        struct inet_bind_bucket *tb;
-       int ret;
+       int ret, attempts = 5;
        struct net *net = sock_net(sk);
+       int smallest_size = -1, smallest_rover;
 
        local_bh_disable();
        if (!snum) {
                int remaining, rover, low, high;
 
+again:
                inet_get_local_port_range(&low, &high);
                remaining = (high - low) + 1;
-               rover = net_random() % remaining + low;
+               smallest_rover = rover = net_random() % remaining + low;
 
+               smallest_size = -1;
                do {
                        head = &hashinfo->bhash[inet_bhashfn(net, rover,
                                        hashinfo->bhash_size)];
                        spin_lock(&head->lock);
                        inet_bind_bucket_for_each(tb, node, &head->chain)
-                               if (ib_net(tb) == net && tb->port == rover)
+                               if (ib_net(tb) == net && tb->port == rover) {
+                                       if (tb->fastreuse > 0 &&
+                                           sk->sk_reuse &&
+                                           sk->sk_state != TCP_LISTEN &&
+                                           (tb->num_owners < smallest_size || smallest_size == -1)) {
+                                               smallest_size = tb->num_owners;
+                                               smallest_rover = rover;
+                                               if (hashinfo->bsockets > (high - low) + 1) {
+                                                       spin_unlock(&head->lock);
+                                                       snum = smallest_rover;
+                                                       goto have_snum;
+                                               }
+                                       }
                                        goto next;
+                               }
                        break;
                next:
                        spin_unlock(&head->lock);
@@ -125,14 +141,19 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
                 * the top level, not from the 'break;' statement.
                 */
                ret = 1;
-               if (remaining <= 0)
+               if (remaining <= 0) {
+                       if (smallest_size != -1) {
+                               snum = smallest_rover;
+                               goto have_snum;
+                       }
                        goto fail;
-
+               }
                /* OK, here is the one we will use.  HEAD is
                 * non-NULL and we hold it's mutex.
                 */
                snum = rover;
        } else {
+have_snum:
                head = &hashinfo->bhash[inet_bhashfn(net, snum,
                                hashinfo->bhash_size)];
                spin_lock(&head->lock);
@@ -145,12 +166,18 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 tb_found:
        if (!hlist_empty(&tb->owners)) {
                if (tb->fastreuse > 0 &&
-                   sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
+                   sk->sk_reuse && sk->sk_state != TCP_LISTEN &&
+                   smallest_size == -1) {
                        goto success;
                } else {
                        ret = 1;
-                       if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb))
+                       if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb)) {
+                               if (sk->sk_reuse && sk->sk_state != TCP_LISTEN && --attempts >= 0) {
+                                       spin_unlock(&head->lock);
+                                       goto again;
+                               }
                                goto fail_unlock;
+                       }
                }
        }
 tb_not_found: